Even the most sophisticated artificial intelligences may be subject to the biases of their creators. At best, these algorithmic biases can lead to inaccurate or nonsensical outputs. At worst, they can cause real and lasting harm..
If artificial intelligence and machine learning are ever to be used responsibly and ethically, we must first find a way to mitigate bias.
There are many different types of bias — but within the context of machine learning, these biases are all caused by one or more of the following issues:
In most cases, these biases are unintentional. Unfortunately, while biases in human decision-making may have small impacts, even a small bias in an algorithm may lead to exponentially worse decision-making. This is because an AI doesn't simply learn the biases it's taught — it tends to amplify them.
Bias mitigation is the process of identifying and eliminating the biases that may be present in an artificial intelligence. This may be accomplished manually or through bias mitigation algorithms. Mitigation techniques can be divided into several broad categories, and may be either manual or algorithm-based.
Manual bias mitigation techniques include:
Unfortunately, none of these techniques are necessarily effective at addressing unconscious biases. For this reason, it's generally advisable to also incorporate algorithmic bias mitigation strategies.
Bias mitigation for machine learning can be divided into three broad categories based on where in the machine learning model's life cycle it is applied.
Techniques such as data preprocessing assess training data for potential biases before it's fed into a machine learning model, with the goal of ensuring said data is as fair and impartial as possible. Unfortunately, this does little to address the biases that may be present in the model itself.
In-processing bias mitigation occurs during training, and generally comes in one of two forms. Fairness-aware algorithms are explicitly designed to be impartial, with checks and balances to ensure that they don't favor or discriminate against any group. Adversarial debiasing, meanwhile, trains the machine learning model alongside an adversary designed to identify and mitigate biased decision-making.
After a model has been trained, post-processing bias mitigation may be applied to ensure it remains impartial. This form of bias mitigation works in one of two ways. Either it leverages probabilities to change outputs, or it identifies groups experiencing bias and projects favorable outcomes onto them.
While this technically accomplishes the objective of mitigating bias, it also impacts the accuracy of the system. As such, it may not necessarily represent an ideal solution to the issue of bias.
